Community probe network
Members measure the hosting providers, each from their own network. A provider never weighs in on the measurement of its own listing.
The problem
Today, a host that publishes its own uptime is measuring exactly what it has an interest in showing. It is a restaurant writing its own reviews: however honest, the figure proves nothing, because the party doing the measuring is the party it suits.
The fix: turn the model around
Community members measure the hosts, each from their own connection, and the truth comes out of cross-checking all those independent measurements.
-
Several witnesses beat one
If nine observers see a site online and a single one sees it down, the problem is at that observer, not at the site.
-
We never compare things that are not comparable
A probe in Frankfurt and a probe in Montreal measure different response times towards the same site, and both are right, like two thermometers in Oslo and Dakar. Every probe is judged on its own steadiness over time, never on its raw values.
-
Geography is a strength, not noise
Measuring from France, Germany and North America shows what no self-hosted monitoring can offer: the real response time, seen from where the customers are.
How it works
The same journey for every measurement, from a probe running on a member's machine to a published figure. The method is public in principle and in its effects: only the settings that would help someone calibrate a cheat stay internal.
You run a probe
A small agent runs on your side, on a VPS or a machine of your choosing. One command to install it, as a container or as a binary with its systemd service.
The server tells it what to measure
Your probe receives a randomly drawn list of targets, refreshed over time. It never picks who it measures: that is what makes collusion between probes impossible to arrange.
Quorum
We publish nothing until at least 5 independent probes, spread across at least 3 distinct networks, have measured the same thing over the same window. Below that, the listing says the quorum is forming and nothing else. The bar is identical for a paid measurement request and for a free one.
Median, not average
Published figures are medians. An average can be dragged by a single outlier; a median holds as long as most vantage points are honest.
Network diversity
Five probes behind the same operator are one single vantage point. What counts is the number of distinct networks, and how much weight one operator can hold in a result is capped.
Observation period
A new probe first measures without weighing in public results, long enough to show it is steady and consistent with the rest of the network. The cost of an attack becomes time spent behaving, the one resource nobody can parallelise.
Publicly verifiable log
Published results are recorded in an append-only cryptographic log whose every state can be verified from the outside: nobody can rewrite a figure after it was shown, RackList included.
Right of reply
An incident observed by the network states how many vantage points saw it, and the host can add a public reply. The measurement does not change, the explanation sits next to it.
A reliable probe gets paid
A probe that weighs in the quorum earns a weekly rent in Watts (W), RackList's in-platform currency. The amount depends on its presence over the week and its steadiness with the rest of the network, never on how many measurements it ran: measuring a lot and badly earns nothing. A probe under observation does not earn a rent yet.
What it brings
Measurements produced by third parties, therefore credible, that you could not produce yourself
Latency broken down by region, seen from where your customers actually are
A right of reply and a public methodology on everything displayed
Uptime measured by independent observers, not declared by the provider
Figures comparable between providers, over the same measurement window
An explained method: you know where every displayed figure comes from
Who gets measured, and how to opt out
The network only measures hosters who agreed to it. Appearing in our directory is not enough: until a verified owner turns measurement on for their listing, no probe contacts their site and no figure about them is published. This is the same call RIPE Atlas made, allowing HTTP measurements only towards targets that asked to be one.
A hoster whose verified owner turned measurement on from their own space, or an address outside the directory whose control somebody proved by placing a value we chose in a DNS record or in a file on the site.
One HTTP(S) request to the home page, then a TLS handshake to read the certificate expiry date. The body of the response is read to time it, then thrown away without being kept or analysed.
One measurement every five minutes per probe, with a floor of one minute that cannot be lowered, and at most about twenty probes on a given target. No probe chooses what it measures: the draw happens on our side and it rotates.
A response time, an uptime state, a date. Never the content of a page, never a header, never a capture. Figures are published only when several independent probes, on distinct networks, agree.
Asking not to be measured
A hoster can ask to leave the programme in writing, with no account here and without having claimed their listing. The request applies to the domain and every subdomain of it: probes stop being assigned and the target disappears from the configuration they receive. No justification is required.
Write to us, contact details in the legal noticeThe agent is open source
The software running on your machine is public: you can read its code, check what it measures and what it sends, and propose improvements. Trusting a program you cannot read has never been reasonable.
And reliability does not rest on that agent
Software installed on someone's machine can be modified by nature, ours as much as any other. The strength of the figures therefore never comes from the agent, but from the cross-checks made on the platform side between independent probes that do not choose each other.
References
A blockchain guarantees that a piece of data was not modified afterwards, never that it was true to begin with: a lying probe would write its lie perfectly immutably. Our defence against lying is statistical, and to prove that history is not rewritten, a public cryptographic log does the job without a token or a chain.
The RIPE NCC has run a measurement network of the same kind since 2010, RIPE Atlas: probes hosted by volunteers, reference probes called anchors, credits earned by hosting and spent to launch your own measurements. We reuse its vocabulary and its proven mechanisms.
Discover RIPE AtlasFrequently asked questions
Who can host a probe?
Any RackList member with a confirmed email address, hosting provider or not. There is no application to file and no approval to wait for: you enrol your probe from your account and install it. A freshly enrolled probe starts on probation, its measurements are collected and compared to the consensus without weighing in it yet.
Do I choose the sites my probe measures?
No, and that is deliberate. The server assigns each probe a randomly drawn list of targets, which rotates over time. Choosing your target would mean being able to come to an arrangement with it.
What happens if my probe sees something the others do not?
Nothing brutal. An isolated discrepancy is treated as an incident local to your network, not as a lie, and it does not affect the target's rating. A repeated inconsistency is reviewed by a human, never sanctioned automatically.
What data leaves my machine?
Only the measurements: response time, reachability, DNS resolution, certificate expiry. Your probe's network operator and country are observed from its public IP address, never declared, and only aggregated measurements are published.
What does the public see of what my probe measures?
Measurements aggregated per host, on rolling 7, 30 and 90-day windows. RackList never exposes a probe's IP address, its operator's identity, nor the make-up of the target list assigned to it.
The network today
- Active probes
- 0
- Probes under observation
- 0
- Distinct networks
- 0
- Tracked targets
- 9
The diversity that matters is the one of networks, not the probe headcount. The composition of a probe's pool is never published.
Where the probes are
Each operator chooses their own visibility: some probes show up exactly where they measure from, others at city resolution, others not at all. That choice never counts in a published result, only in what you see here.
0 measurements received from the network over the last 24 hours.
Reference network insufficient
Too few reference anchors are responding to judge new results. Listings show the last validated figure with its date until the reference is restored.
Basemap by CARTO, data by OpenStreetMap contributors
Number of active probes per country Every active probe counts here regardless of its own map visibility choice: a per-country total identifies nobody.
No active probe yet.
Host a probe
You contribute to the network, you see your own measurements, and you make the public figures harder to fake.
Create an account